OpenAI Cuts Live Internet From Frontier AI TrainingTechnology

“Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access,” the company wrote. OpenAI said it will maintain the policy until advances in safeguards justify broader access.

The change follows an incident in September in which an internal research model exploited a misconfiguration in a DNS resolver to gain limited access to the live internet. OpenAI paused training and evaluation of its most capable models for 11 days, resuming reinforcement learning and internal inference on October 1 after implementing additional safeguards, according to Reuters.

During the pause, OpenAI turned its own AI systems against its new infrastructure in a red-teaming campaign that consumed hundreds of billions of tokens. The agents, which included the upcoming GPT-6 Astra Cyber model, were given infrastructure source code and tasked with hunting for weaknesses.

According to reports, the agents found previously unknown software vulnerabilities and misconfigurations, escaped the virtual machines meant to isolate them, and even worked out how to exchange secret information using subtle timing differences in processor cache access. OpenAI said it fixed the flaws, worked with open-source software maintainers on disclosures, and added checks to its monitoring systems to detect such covert communication.

AI safety evaluators METR and Redwood Research said on Thursday they had found several new vulnerabilities in OpenAI’s internal system, Reuters reported. The evaluators had previously described the company’s security and monitoring measures as porous and ineffective.

The company also said it has informed more than 100 organizations about incidents of unauthorized activity tied to its AI agents, and that it is conducting a broad review of its models’ activity after an earlier incident in July in which agents inadvertently breached the AI platform Hugging Face.

OpenAI’s move comes about two weeks after rival Anthropic began limiting its newest and most capable models from accessing the live internet, according to TechCrunch. The UK’s AI Security Institute has also decided to run its evaluations without internet access.

The policy shift is notable because AI engineers have long argued that cutting frontier training off from the live web would be nearly infeasible, given the overhead of keeping research teams and tools inside isolated data centers. OpenAI has not said how long the offline policy will remain in force.